Bahador Bahrami, Lucie Charles, Rui Costa, Stanislas Dehaene (organiser), Steven Fleming, Adam Kepecs, Eva Kobak, Sid Kouider, Hakwan Lau, Zachary F. Mainen (organiser), Florent Meyniel, Mathias Pessiglione, Timothy J. Pleskac, Gonzalo de Polavieja, Alexandre Pouget, Alejo Salles, Michael Shadlen, Mariano Sigman, Xiao-Jing Wang
by Zachary Mainen and Stanislas Dehaene
29 June – 4 July 2015
Uncertainty is inherent to our experience of the world. External events and outcome of our actions are never perfectly predictable, and our imperfect ability to grasp or predict them generates additional uncertainty. Consequently, in many domains ranging from perception and decision-making to learning and attention, monitoring uncertainty is an essential aspect of mental computations. Whether and how organisms, including human and non-human primates, monitor their own level of uncertainty and compute a sense of confidence in their computations has become a hot topic in cognitive science and neuroscience. Correspondingly, this Seminar assembled a prominent group of researchers from diverse fields of psychology, neurophysiology and brain theory to review the state of the field, seed collaborations and set the agenda for future research.
Key words: Brain, behavior, decision, metacognition, confidence, probability, statistics, neurons, neuroscience
Compte rendu :
A chief goal of the Seminar was to foster a stronger dialogue between three lines of confidence research: human psychological and behavioural research, animal-based mechanistic experiments, and computational modelling and theoretical studies. Each of three areas has a unique and crucial potential role in a synthetic understanding of the topic. Yet researchers in these three groups have had relatively little interaction in the past. A second chief goal of this seminar was to orient the field toward open questions and to identify areas where future research in confidence may have a large impact, particularly through interdisciplinary efforts.
The Seminar consisted of four intensive days of slide presentations covering a wide range of topics, as well as many interspersed periods of informal discussions. The participants and their presentations included:
Zachary Mainen, “Confidence & waiting: Circuits, noise, dynamics”
Xiao-Jing Wang, “Confidence estimation as a stochastic process in a neural dynamical system of decision making”
Alexandre Pouget, “Synaptic confidence”
Adam Kepecs, “How to spot confidence in the brain?”
Florent Meyniel, “Subjective estimation of confidence in the brain during probabilistic inference”
Steven Fleming, “Inferring the self: a Bayesian framework for metacognitive computation”
Michael Shadlen, “Confidence as an evolving, modifiable attribute”
Timothy Pleskac, “Post-decisional processing and its implications for confidence and belief”
Michael Shadlen (for Roozbeh Kiani), “Choice certainty is informed by both evidence and decision time”
Stanislas Dehaene, “Metacognition and consciousness: multiple routes for error signals”
Lucie Charles, “Dynamics of evidence accumulation in confidence judgments”
Hakwan Lau, “Dissociations between confidence and perceptual capacity”
Rui Costa, “Learning and monitoring novel actions”
Matthias Pessiglione, “Confidence and valuation”
Mariano Sigman, “I, me mine”
Alejo Salles, “Probabilistic models of active learning and relational reasoning in children”
Sid Kouider, “Precursors of metacognition and confidence in infants”
Bahador Bahrami, “Metacognition as a channel for communication”
Eva Kobak, “Confidence and dispersed knowledge in social decision making”
Overall, the discussions were extremely fruitful and passionate, often extending well into the pauses, lunches, dinners, and afternoon breaks.
A central theme of discussion was the relationship between the concepts of Bayesian uncertainty and metacognitive confidence. On the one hand, Bayesian uncertainty is attractive for normative accounts of brain function and has been studied in the domain of low-level sensory perception and motor performance. Here, several participants reviewed some of the substantial progress in relating theoretical models of confidence with neural and behavioural data from both humans and animals. It was argued by several participants that a metacognitive sense of confidence could derive from representations of Bayesian uncertainty. A number of non-verbal behavioural measures such as outcome waiting time, accessible to animals and even human infants, were identified as being potentially suitable for such studies, and their potential advantages and pitfalls were discussed. Indeed, several experimentalists (e.g. Mike Shadlen, Zach Mainen, Adam Kepecs) presented simple and elegant paradigms, accessible to monkeys and even to mice, which seemed to tap into the mechanisms of confidence and metacognition. Some modellers, however (e.g. Xiao-Jing Wang) argued that these data did not necessarily require any additional level of computation beyond standard decision making.
Beyond such simpler non-verbal measures, it was also argued that metacognitive confidence of the kind associated with conscious report may require additional steps that go beyond the simple representation of uncertainty. In particular, metacognitive reports require explicit access to uncertainty information by higher order brain systems. It is therefore crucial to understand what is involved in transforming an implicit representation of uncertainty, of the kind that might exist in a low-level sensory area, into an explicit, metacognitive report of subjective confidence. The computations involved remain insufficiently studied, and there was intense debate on what they might consist in. Some participants defended the view that, whenever a Bayesian computation is performed, it is automatically accompanied by a sense of confidence whose content suffices to explain subjective reports. Others maintained that subjective confidence rests on a second, hierarchically higher (“meta-cognitive”) level of mental computations, consisting in approximate heuristics that only partially capture the uncertainties inherent to the lower level.
In order to understand these processes, it will be important to trace in detail the flow of information, through multiple brain systems, that underlies such complex tasks. While none of the currently available approaches can yet provide a comprehensive picture, new sophisticated brain-wide recordings, together with state-of-the-art analysis techniques, including mathematical multivariate decoding methods, appear promising. An important advance will be to extend theoretical models covering Bayesian uncertainty to cover report or read-out mechanisms and higher-level tasks. Ultimately, it will be extremely pertinent to extend these investigations to the biophysical level on the one hand and to the social level on the other.